GDC 2015 AI Tricks session

 
March 2015 – The opening day of the GDC conference was very crowded, and was a strong precursor for the record attendance setting event this year. With over 26,000 registered attendees for the first time, most of the developers and coding sessions were shifted to rooms seating over 300 people at the main convention center, and over 200 people in the expansion “West” hall building.

In a session that is always well attended, the almost 500 people in room heard from several developers who worked on globally recognized titles, on what were some of their favorite tricks on implementing in game AI techniques. Most of these tricks are detailed on the sites gameai.com and gameaipro.com .

Two of the tricks stood out as being good general information, regardless of type of game being developed and the skill level of the developer. The first was related to “how fast should the AI react” and the second is “understanding the context of the AI”.

When most developers encounter AI control of characters in the game, one of the major decisions they need to make is “how fast should I react?” The gut feel of the designer is “quickly”, but in computer terms that is kind of nebulus. The results of working on many games and customer feedback on these games is a good starting place for the optimization is 0.25 sec (250ms) for an action based decision and 0.4 sec (400ms) for a go/no go type of decision for the character. This allows for smooth continutity of play, while not frustrating the user that the “game is not reactive”.

A second tip came from one of the developers of the Sims. This game has very sophisticated AI on multiple levels and dynamic interacting code. As a result, the developer can get a false sense of the control for the reaction of the character as it may not be obvious which AI engine is leading. He suggested a very humbling experience of taking a break and programming a robot. He brought to the room a very simple Arduino Uno based robot that had IR detectors, wire “bump” detectors and simple controls for forward, reverse, right, left control. This is a “single input & response” system. In the course of planning and playing with the robot, he “rediscovered” how simplistic, directed and context unaware the sensor response is. This allowed him to refocus on the “main action” defining input and response.